Hugging Face's logo

jaeyong2
/
gte-multilingual-base-Ko-embedding

Feature Extraction
Transformers
Safetensors
Korean
new
custom_code
text-embeddings-inference
Inference Endpoints
Model card Files Community
gte-multilingual-base-Ko-embedding / README.md
jaeyong2's picture
jaeyong2
Update README.md
190ce84 verified
preview code
|
raw
history blame contribute delete
3.82 kB
metadata
library_name: transformers
license: apache-2.0
datasets:
  - jaeyong2/Ko-emb-PreView
language:
  - ko
base_model:
  - Alibaba-NLP/gte-multilingual-base

Model Card for Model ID

Model Details

Train

  • H/W : colab A100 40GB
  • Data : jaeyong2/Ko-emb-PreView
model_name = "Alibaba-NLP/gte-multilingual-base"
dataset = datasets.load_dataset("jaeyong2/Ko-emb-PreView")
train_dataloader = DataLoader(dataset['train'], batch_size=8, shuffle=True)

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name).to(torch.bfloat16)
triplet_loss = TripletLoss(margin=1.0)

optimizer = AdamW(model.parameters(), lr=5e-5)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

for epoch in range(3):  # 에포크 반복
    model.train()
    total_loss = 0
    count = 0
    for batch in tqdm(train_dataloader):
        optimizer.zero_grad()
        loss = None
        for index in range(len(batch["context"])):
            anchor_encodings = tokenizer([batch["context"][index]], truncation=True, padding="max_length", max_length=4096, return_tensors="pt")
            positive_encodings = tokenizer([batch["Title"][index]], truncation=True, padding="max_length", max_length=256, return_tensors="pt")
            negative_encodings = tokenizer([batch["Fake Title"][index]], truncation=True, padding="max_length", max_length=256, return_tensors="pt")

            anchor_encodings = batch_to_device(anchor_encodings, device)
            positive_encodings = batch_to_device(positive_encodings, device)
            negative_encodings = batch_to_device(negative_encodings, device)

            # 모델 출력 (임베딩 벡터 생성)
            anchor_output = model(**anchor_encodings)[0][:, 0, :]  # [CLS] 토큰의 벡터
            positive_output = model(**positive_encodings)[0][:, 0, :]
            negative_output = model(**negative_encodings)[0][:, 0, :]
            # 삼중항 손실 계산
            if loss==None:
                loss = triplet_loss(anchor_output, positive_output, negative_output)
            else:
                loss += triplet_loss(anchor_output, positive_output, negative_output)
        loss /= len(batch["context"])
        loss.backward()
        optimizer.step()

Evaluation

Code : 

import torch
import numpy as np
from sklearn.metrics import pairwise_distances
from tqdm import tqdm


dataset = datasets.load_dataset("jaeyong2/Ko-emb-PreView")
validation_dataset = dataset["test"].select(range((1000)))

model.eval()

def evaluate(validation_dataset):
    correct_count = 0

    for item in tqdm(validation_dataset):
        query_embedding = get_embedding(item["context"], model, tokenizer)
        document_embedding = get_embedding(item["Title"], model, tokenizer)
        negative_embedding = get_embedding(item["Fake Title"], model, tokenizer)
      

        # 쿼리와 모든 문서 간의 유사도 계산 (코사인 거리 사용)
        positive_distances = pairwise_distances(query_embedding.detach().cpu().float().numpy(), document_embedding.detach().cpu().float().numpy(), metric="cosine")
        negative_distances = pairwise_distances(query_embedding.detach().cpu().float().numpy(), negative_embedding.detach().cpu().float().numpy(), metric="cosine")

        if positive_distances < negative_distances:
            correct_count += 1

    accuracy = correct_count / len(validation_dataset)
    return accuracy

results = evaluate(validation_dataset)
print(f"Validation Results: {results}")

Accuracy

  • Alibaba-NLP/gte-multilingual-base : 0.974
  • jaeyong2/gte-multilingual-base-Ko-embedding : 0.981

License